Expanding commonsense knowledge bases by learning from image tags

Commonsense knowledge;transfer learning;image collection;knowledge extraction

Mauceri, Cecilia R ; Lazebnik ; Svetlana

关键词: Commonsense knowledge;    transfer learning;    image collection;    knowledge extraction;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/88091/MAUCERI-THESIS-2015.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

I present a method for learning new commonsense facts to augment existing commonsense knowledge bases by using the metadata of large online image collections. Online image collections present a source of knowledge that is supported by many contributors, has good representation of objects and their properties, and is visual. The collection's broad support of objects and object properties ensure the relevance and quality of the commonsense knowledge collected, while the visual focus provides a different subset of knowledge than typical text corpora. Using the image metadata provides a text representation of the visual information. Therefore, I can use classifiers trained on existing text-based knowledge bases to learn relationships between concepts represented in the images. I collect two datasets of more than 1 million images each, one consisting of animal images, one of room interiors. The images are tagged with relevant concepts by their owners. I train classifiers using facts from two popular commonsense knowledge bases, ConceptNet and Freebase, to classify the relationships between frequent concept pairs. The output is a list of more than 90,000 proposed facts, which are in neither source knowledge base.